Improving Long Tailed Document-Level Relation Extraction via Easy Relation Augmentation and Contrastive Learning

Abstract

Towards real-world information extraction scenario, research of relation extraction is advancing to document-level relation extraction(DocRE). Existing approaches for DocRE aim to extract relation by encoding various information sources in the long context by novel model architectures. However, the inherent long-tailed distribution problem of DocRE is overlooked by prior work. We argue that mitigating the long-tailed distribution problem is crucial for DocRE in the real-world scenario. Motivated by the long-tailed distribution problem, we propose an Easy Relation Augmentation(ERA) method for improving DocRE by enhancing the performance of tailed relations. In addition, we further propose a novel contrastive learning framework based on our ERA, i.e., ERACL, which can further improve the model performance on tailed relations and achieve competitive overall DocRE performance compared to the state-of-arts.

PRiSM: Enhancing Low-Resource Document-Level Relation Extraction with Relation-Aware Score Calibration

“…With the introduction of DocRED (Yao et al, 2019), many approaches were proposed to extract relations from a document (Wang et al, 2019;Ye et al, 2020;Zhang et al, 2021;Xu et al, 2021;Zhou et al, 2021;Xie et al, 2022). The long-tailed data problem of DocRE has been addressed in some studies (Du et al, 2022;Tan et al, 2022a), as well as low-resource DocRE (Zhou et al, 2023); however, most require additional pretraining, which is compute-and cost-intensive, while PRiSM only requires adjusting logits in existing models. Lowresource RE has been extensively studied at the sentence level, and we specifically focus on leveraging label information (Yang et al, 2020;Dong et al, 2021;Zhang and Lu, 2022) in which PRiSM applies it to the document level.…”

“…We argue that the reason is twofold. First, the long-tailed distribution of DocRE data encourages models to be overly confident in predicting frequent relations and less sure about infrequent ones (Du et al, 2022;Tan et al, 2022a). Out of the 96 relations in DocRED (Yao et al, 2019), a widely-used DocRE dataset, the 7 most frequent relations account for 55% of the total relation triples.…”

“…PRiSM then learns re-lation representations to output adaptive scores for each relation triple. Note that previous work mostly utilized relation representations for self-supervised learning (Dong et al, 2021;Du et al, 2022;Zhou et al, 2023), whereas PRiSM uses them to directly adjust logits, which brings a calibration effect. To elaborate further, let us say that classification logits are statistical scores and similarities are semantic scores.…”